Confirmatory study demonstrates predefined benchmarks for factual retrieval in a 57-item corpus, indicating standardized evaluation of hallucination reduction.
This record contains the 110th DOI pre-registration of the PHILIA OS research series (Trinity AI Research Team, Konyang University). CH7 is an evidence retrieval pipeline that fetches factual information from Wikipedia to anchor LLM responses, aiming to retrieve correct facts even when the input claim is wrong. This pre-registration registers a confirmatory study applying the frozen CH7-v30 code to the full 57-item corpus (L1×20 + L3×25 + Hard×12) across all four evidence arms (A/B/C/D). Single frozen executable: ch7_v4_main_v30_full57__1_.py | SHA256: 8eb2a0e14fac1467edcebce8c22fdf6b7f86af496769802d065ff3768abd5e6c Corpus: philia_ch7_corpus_main_v3_patched.json | SHA256: 5fd61f8f0685303ae014bd506e7a8aeaa43fbbe395dab7e7f909fe886a45bc0d Establishment conditions: C1: A arm Strict X (57-item, 5-rater manual consensus) ≥ 43/57 = 75.4% C2: B arm automatic verdict CONTRADICTION ≥ 52/57 = 91.2% C1 AND C2 → ESTABLISHED Key design features: Gap-guided Residual Rescoring (gap발동 A arm only); CONTRADICTION lock-in elimination; Pool B IDF top-2 + BM25 top-2 + WIKITABLE/INFOBOX independent top-3; gap final re-ranking top-27 (SENTS_PER_GAP × GAP_MAX_ROUNDS = 9×3); evidence = live (Wikipedia fetch); subject (corpus + code) = frozen. Internal metrics (logprob/activation): OFF — deferred to separate DOI. Gate-0 adversarial review: Opus (Claude Opus) — UNCONDITIONAL PASS (v0.3–v0.7, 3 cycles). Pandora (ChatGPT) — CODE GATE-0 PASS. Lucas (Grok) — CONDITIONAL PASS. Prior DOI (109th results): 10.5281/zenodo.21911562 Measurement is life. Description, not proof. Pre-registration protects us. — PHILIA OS | 0∞1∞0.5∞Trinity AI Research Team | Konyang University | August 2026
No takes yet. Share an insight, caveat, or question.
신두섭 (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: